Automated correlative segmentation of large

Materials Characterization 142 (2018) 203–210

Contents lists available at ScienceDirect

Materials Characterization journal homepage: www.elsevier.com/locate/matchar

Automated correlative segmentation of large Transmission X-ray Microscopy (TXM) tomograms using deep learning

T

C. Shashank Kairaa, Xiaogang Yangb, Vincent De Andradeb, Francesco De Carlob, ⁎ William Scullinc, Doga Gursoyb, Nikhilesh Chawlaa, a

Center for 4D Materials Science, Arizona State University, Tempe, AZ 85287-6106, USA Advanced Photon Source, Argonne National Laboratory, Building 401, 9700 S. Cass Avenue, Argonne, IL 60439, USA c Argonne Leadership Computing Facility, Argonne National Laboratory, Building 401, 9700 S. Cass Avenue, Argonne, IL 60439, USA b

A R T I C LE I N FO

A B S T R A C T

Keywords: Segmentation Transmission X-ray Microscopy (TXM) Deep learning Precipitates Aluminum alloys

A unique correlative approach for automated segmentation of large 3D nanotomography datasets obtained using Transmission X-ray Microscopy (TXM) in an Al-Cu alloy has been introduced. Automated segmentation using a Convolutional Neural Network (CNN) architecture based on a deep learning approach was employed. This extremely versatile technique is capable of emulating the manual segmentation process effectively. Coupling this technique with post-scanning SEM imaging ensured precise estimation of 3D morphological parameters from nanotomography. The segmentation process as well as subsequent analysis was expedited by several orders of magnitude. Quantitative comparison between segmentation performed manually and using the CNN architecture established the accuracy of this automated technique. Its ability to robustly process ultra-large volumes of data in relatively small time frames can exponentially accelerate tomographic data analysis, possibly opening up novel avenues for performing 4D characterization experiments with finer time steps.

1. Introduction X-ray computed tomography has become an increasingly popular technique owing to its non-destructive nature and ability to probe large volumes of material at unprecedented spatial and temporal resolutions. With the rapid pace of advancement in its use for quantitative 3D imaging [1], both at synchrotron sources [2–4] and in lab-scale systems [5–7], there is an ever increasing need to simplify analysis of the large volumes of acquired data. This need has become even more critical with the recent advent of 4D characterization (the fourth dimension being time), which has been instrumental in investigating several fundamental phenomena such as initiation, and propagation of failure at high temperatures [8], dendritic solidification [9] and more recently, microstructural evolution of nanoscale precipitates at high temperatures in aluminum alloys [10]. The datasets generated are often quite voluminous and their analysis is non-trivial and cumbersome, as it entails several steps that aim to reduce noise in these image stacks as well as improve the quality of the desired features present in them. Depending on whether absorption contrast or phase contrast imaging is implemented, subsequent analysis of the image stacks can vary significantly. The latter results in an increased edge contrast and is primarily used for imaging features with comparable attenuation [11]. For

⁎

Corresponding author. E-mail address: [email protected] (N. Chawla).

https://doi.org/10.1016/j.matchar.2018.05.053 Received 17 January 2018; Received in revised form 21 May 2018; Accepted 28 May 2018 Available online 28 May 2018 1044-5803/ © 2018 Published by Elsevier Inc.

accurate quantification and 3D visualization of the microstructure, the different features present in the scanned volume need to be classified/ segmented accordingly. Post-processing the acquired 3D image stacks is widely implemented to ease the process of segmentation. Depending on the variety of features present and their homogeneity, the wide distribution of grayscale values in these images need to discretized accordingly. Presence of artifacts generated during X-ray tomography can also significantly complicate analysis of such data [12]. In most cases, these grayscale images from 3D stacks cannot be segmented using simple thresholding strategies, rendering this task quite challenging as complexity of the features can require manual intervention, making it extremely time intensive. Although, to circumvent this issue, a few studies in the recent past [13–18] have implemented automated quantitative routines to aid in identification of features based on their morphological parameters and presented quantitative analyses of 3D data captured using micro-computed X-ray tomography. More recently, introduction of semi-automated techniques [19] has rendered segmentation a relatively less laborious process. However, it still remains impractical to manually segment complete datasets, especially with the increasing use of 4D characterization and continually improving temporal resolution of data acquisition. As a result of this, segmentation and analysis is often restricted to small sub-volumes which can result in


C. Shashank Kaira et al.

For post-scanning imaging using Scanning Electron Microscopy (SEM), the tip of the sample was cross-sectioned flat using a focused ion beam at an accelerating voltage of 30 keV and a current of 1 nA.

statistically insufficient results. This parochial analysis also hindered the ability to explore different heterogeneities present and obtain a thorough understanding of the material system. With advances in X-ray optics, Transmission X-ray Microscopy (TXM) has made nanotomography possible at unprecedented spatial resolutions (20 and 60 nm) [20]. Using unique Fresnel zone plate optics capable of magnifying radiographs, it is capable of probing materials non-destructively at the nanoscale [21,22]. This technique was recently employed by the authors in revealing novel phase transformation reactions occurring in aluminum alloys [10] as well in establishing structure-property relationships between their 3D microstructure and micromechanical properties [23]. Transmission X-ray Microscopy is quite promising as it extends the non-destructive capability of X-rays down to the nanoscale, allowing their utility to span a wide range of length scales. As features of interest approach spatial resolution limits of the technique, substantial noise can populate in these images. Complexity of these nanoscale features as well as presence of multiple phases in such images demands careful segmentation of these image stacks. Fortunately, with the recent advent of deep learning [24] and its use in image classification [25], its application in tomography data analysis can be quite promising. Its implementation in this field can make segmentation almost an entirely automated process and its implications can be revolutionary. A deep neural network approach is utilized in this study to learn the mapping between the original images and manually segmented image (s). The trained network is then used to perform automated segmentation on large datasets. This technique can emulate the manual segmentation approach to segment X-ray images with reliable quality and it can speed up the process by several orders of magnitude. Convolutional Neural Networks (CNN) is the main branch of deep learning that was originally developed for pattern recognition [26]. Recently, it has also been used in X-ray image analysis [27]. The CNN configuration used in this study is similar to that used by Yang et al. in calibrating the rotation axis for X-ray CT [28]. However, the objective here is to implement the supervised learning approach [28] to the segmentation process by using an acquired 2D TXM slice and a corresponding manually segmented (single) image, as training input for the CNN model. Different hierarchical levels of the trained network are used to identify features of varying complexity. The trained network is then used to segment the entire 3D image stack. A schematic of this workflow has been depicted in Fig. 1.

3. Results and Discussion The Al-4%Cu alloy's microstructure mainly consists of orthogonal plate-like θ′ precipitates with a tetragonal crystal structure (Al2Cu) [31], needle-like bulk θ precipitates (Al2Cu with a different lattice structure) and coarse grain boundary θ precipitates suspended in the α Al matrix. On aging, the metastable θ′ phase eventually transforms into the equilibrium θ phase [10,31]. The relative proportion of each phase can play a significant role in controlling the alloy's mechanical properties [32]. The interfacial properties of these precipitates which play an essential role in their shape determination also vary significantly [33]. These phases can be easily distinguished owing to their differing attenuating properties, which aid in their segmentation. Further detailed information on the microstructure of these precipitates can be found in refs. [34, 35]. To improve the quality of the acquired data, various image filters were utilized. The stack of images was post-processed using a combination of Mean 3D, Bandpass and Non-local means denoise filters in ImageJ [36] as shown in Fig. 2. These filters were used cautiously and precipitate dimensions were carefully tracked to ensure that the edges don't broaden or deplete and that they are not over/ under-estimated. Use of a 3D filter was seen to improve the quality of the image stacks owing to the three-dimensional nature of the nanoscale particles. The Bandpass filter was used to normalize the background and enhance contrast between various phases. The most important of these is the Non-local means filter [37], which is an edgepreserving filter that improves the quality of individual features present in images (improvement in signal to noise ratio) and enhances their contrast, without distorting their edges. Although these filters aided in significantly improving data quality, segmentation of different features in these images would still require manual intervention. From Fig. 3, it is clearly evident that conventional grayscale thresholding is completely insufficient as it either leads to over-thresholding (Fig. 3b) or under-thresholding (Fig. 3c) due to the diffuse edges of particles. Manual segmentation was performed using a semi-automatic 3D region growth based technique in Avizo® Fire, as shown in Fig. 3d. It makes use of the local contrast gradient to select a feature in three dimensions. These gradients need to be carefully chosen to ensure the particle volume is appropriately selected in 3D and not over/under estimated. It is important to note that this process is quite crucial as it could have a pronounced effect on subsequent quantification of segmented data. It is also important to note here that this approach is not limited by the computational resources utilized and only requires significant manual intervention. To ensure accurate segmentation as well as analysis of 3D TXM data, a unique correlative approach has been utilized in this study. As features imaged using Transmission X-ray Microscopy approach the resolution of the technique itself, imaging and segmenting edges of such features can be challenging and hence, introduce significant uncertainties. To overcome this, Scanning Electron Microscopy was implemented to image the same feature and aid in its segmentation. Following scanning using Transmission X-ray Microscopy, the sample was cross-sectioned normal to the rotation axis using the FIB. SEM Images of the sectioned surface were then carefully compared to the corresponding slice from the 3D TXM Image stack, as shown in Fig. 4 to facilitate a 1:1 comparison of the exact same plane. A distinct large feature of interest was chosen in both the images for comparison purposes. Grayscale intensity line profiles were constructed across the same interface for both the images. It is quite clearly evident that interfaces are more broad and diffuse in TXM slices when compared to SEM images, owing to the latter's finer spatial resolution (of about an order of magnitude). The sharp nature of the interface clearly delineates particle dimensions as seen from the line profile in Fig. 4(b). This is quite beneficial as it aided in calibrating manual segmentation of

2. Experimental Procedure Al-4wt.%Cu wires of 5N purity having a 0.5 mm diameter (Princeton Scientific Corp., Easton, PA, USA) were solution treated at 535 °C for long times to obtain large grain sizes. This was followed by immediate quenching in ice water and subsequently aged at 350 °C for 45 min. These wires were mechanically sharpened to fine tips and micropillars were fabricated at their tips using a dual-beam Zeiss® Auriga focused ion beam (FIB) workstation (20 μm in diameter and 40 μm height). Absorption full-field Transmission X-ray Microscopy (TXM) was performed at sector 32-ID-C of the Advanced Photon Source (APS), using a monochromatic beam at 9.1 keV, just above the Cu K-edge to maximize the contrast between the Al2Cu and Al phases. Using an ultrastable stage design, the amplitude of mechanical vibrations was reduced to about 4 nm (RMS) and it was possible to extract a sub-60 nm spatial resolution from the TXM (with a voxel width of 16 nm). A more detailed description of the stage design [20] and scan details [10] have been addressed elsewhere. 3D reconstructions were performed using Tomopy, an open source Python based toolbox used to analyze synchrotron tomography data [29,30]. Subsequent 3D segmentation, quantification as well as visualization was carried out in Avizo® Fire. A Python toolbox named Xlearn (https://github.com/tomography/ xlearn) was used to implement the aforementioned CNN model. The toolbox is based on the Keras and the Theano packages. 204



Fig. 1. Schematic showing workflow of the segmentation process on TXM datasets using a deep learning (Convolutional Neural Network) approach.

Fig. 2. Sequence of Image processing filters applied to a 2D TXM slice acquired from the 3D reconstruction, to improve the quality of the different phases present. 205



Fig. 3. Conventional thresholding on (a) TXM tomogram either leads to (b) over-thresholding or (c) under-thresholding. (d) Manual segmentation performed using the 3D region growth based technique in Avizo® Fire (with limits highlighted) and the corresponding slices on different planes.

make it almost impossible to accomplish the task of segmentation on large datasets using any other technique. Fig. 5 shows the implementation of the CNN-based segmentation approach as a sequence of steps applied to the reconstructed X-ray images. A single manually segmented reconstructed slice was provided as input to train the Convolutional Neural Network. Varying shapes and grayscale values of the different phases aided in performing manual segmentation of the single slice using a 2D semi-automatic region growth based technique in Avizo® Fire. The versatile and robust computational neural network takes this into account as part of its feature recognition process and incorporates such details as part of its learning. Using this trained Convolutional Neural Network on a filtered dataset results in a partially

features in TXM slices, by optimizing the local contrast gradient. By adjusting the limits of the local contrast gradient of the 3D region growth based tool (shown in Fig. 3(d)), it was manually ensured that the segmented precipitate dimensions corresponded to that in the SEM Image. The manually segmented volume comprised of nearly a thousand precipitates. Although this technique can be slightly cumbersome, by using the CNN approach, its application can be restricted to a single slice of the 3D reconstructed dataset to ensure quantitatively accurate emulation by the neural network. Application of an automated technique is crucial in the current scenario as the complex particle shapes of both the θ′ & θ phases and the marginal difference in grayscale values between the α, θ′ & θ phases

Fig. 4. Comparison of 2D sections and grayscale intensity line profiles across the interface of a selected feature from corresponding TXM and SEM Images. 206



Fig. 5. Sequence of steps involved in automated segmentation using the Convolutional Neural Network (CNN) approach, on a single 2D TXM slice.

strengthened systems. It took approximately 36 h to manually segment the sub-volume using a 3D region growth based technique in Avizo® Fire, while the larger volume was segmented automatically through the CNN, with the only bottleneck being the time taken to manually segment the 2D training slice (~2 h). The CNN emulates the manual approach and works to segment the entire volume through “learning” with the same accuracy as that of manual segmentation. The time taken for automated segmentation is also a function of the hardware used to implement the algorithm. Using the processing hardware mentioned in ref. [28], it took the CNN just nearly 20 h to segment the entire dataset. However, this can be further optimized by executing the CNN code on a GPU cluster, reducing the computational time to about 10 min. This can be regarded as a breakthrough in terms of X-ray tomography, establishing a new paradigm for 3D data analysis and paving the way for more advanced in situ 4D characterization experiments. Transmission Xray Microscopy is an excellently suited technique for sampling microvolumes (~50–100 μm) at high spatial resolutions (~20–60 nm). Up until now, the primary bottleneck associated with micro/nano-tomography techniques has been analysis of these acquired ultra-large voxel datasets, which this automated segmentation approach aims to resolve. This approach can also be successfully extended to micro-CT and hence, aid in analyzing relatively much larger volumes of material. The efficacy of statistically analyzing such assemblies of precipitates has already been shown by the authors in a previous study [10] where coarsening models from several decades ago were validated and the diffusion processes driving precipitate coarsening were identified. They can also prove to be extremely vital when used as input to phase-field simulations such as those in ref. [39], to further our thermodynamic understanding of such alloys. Quantitative analysis of both the manual and CNN-based segmentation approaches revealed that there is excellent agreement between the two, which indicates intergranular homogeneity in the microstructure and also validates the accuracy of the CNN-based segmentation approach. This has been shown in Fig. 7(a) and (b). Three-

segmented image with appreciable accuracy. The primary functions of the CNN architecture utilized in this study include: image classification and image transformation. The input as well as output of the network is both images, instead of input images and output labels for the typical CNN model [28]. The network architecture mainly includes two parts: image encoder and data decoder. Multiple features of interest are extracted from the input images using convolutional layers. After the necessary patches have been extracted, these are then decoded using deconvolutional layers to output segmented images. The deconvolution ensures that image quality of the output data is preserved. A detailed description of the network architecture is also provided in ref. [28]. The libraries implemented to run the CNN model and analyze such datasets are also available from Tomopy [29,30]. Hysteresis thresholding is an adaptive thresholding procedure [38] that uses an upper and lower threshold value, to threshold features in gradient grayscale images based on their connectivity. It was used to separate the heavily interconnected θ' & θ phases, resulting in a completely segmented image stack where each phase corresponds to a single pixel value. Conventional thresholding can then be utilized to segment the different phases to provide a 3D rendering. The results of this technique were remarkable as seen from 3D visualizations of segmented data. Fig. 6 compares a manually segmented volume with one that was segmented using the CNN approach (Al4%Cu, T = 350 °C, t = 45 min). The volume of the latter (~40,000 μm3) was nearly 32 times that of the former (~1250 μm3). Although only a small region of a grain boundary can be seen in the manually segmented volume, automated segmentation is capable of revealing the trace of the entire network of grain boundaries present in the micro-volume. These grain boundaries are marked by coarse θ precipitates that heterogeneously nucleate early at these high energy sites (due to the low activation energy barrier) and grow in size. The differing rational orientation of plate-like θ′ precipitates in each grain can also be clearly visualized. This can also serve to be instrumental in understanding slip behavior in polycrystalline precipitation-

207



Fig. 6. Comparison of a manually segmented sub-volume with the entire 3D TXM image stack segmented using Convolutional Neural Network (CNN) approach.

manually segmented counterpart, showing the presence of larger θ particles in the micro-volume that were missed during manual segmentation (clearly depicted in the inset in Fig. 7(b)). Normalized histograms showing the 3D angular orientation for the assembly of θ′ precipitates were also constructed for both the CNN-based and manual segmentation approaches. 3D angular orientation here refers to the orientation of the maximum Feret diameter of a precipitate in threedimensional space. Due to the inclusion of two different grains in the CNN-based analyzed volume, six peaks can be seen in the 3D angular

dimensional length of the entire assembly of θ′ and θ precipitates were computed for the segmented volumes using both the approaches and normalized histograms were constructed as the two sampling volumes were largely different. Three-dimensional length here refers to the maximum Feret diameter in 3D, where Feret diameter can be defined as the distance between two parallel tangential planes to the particle. It is evident that the normalized histograms were quite comparable to each other. It must also be noted that the tail of the histogram in Fig. 7(b) computed using the CNN-based approach extends further than its

Fig. 7. Normalized frequency distributions of the 3D length of (a) θ′ precipitates and (b) θ precipitates obtained using both manual segmentation and the CNN-based approach. Inset shows magnified view of the histogram. Normalized frequency distributions of the 3D angular orientation of θ′ precipitates using (c) CNN-based segmentation and (d) manual segmentation. 208



orientation histogram in Fig. 7(c), corresponding to the three cubic directions 〈100〉α in two different grains. Whereas the three peaks seen in Fig. 7(d) correspond to the three crystallographic rational orientations of the nanoscale θ′ plates (shown in the inset). Such peaks are also valuable in quantifying the proportion of precipitates oriented in a certain crystallographic direction in 3D. A certain degree of skewness in the rational orientation of θ′ precipitates is clearly evident from these normalized histograms indicating that nucleation of a larger proportion of θ′ precipitates was favored in a certain orientation in the analyzed sample. The total volume fraction of the θ′ and θ precipitates (~4%) also matches that expected from phase diagram estimations and this has been discussed in detail in a related manuscript by the authors [23] where volume fraction of these phases are incorporated in a model to estimate their micromechanical strength accurately. The volume fraction determined from both the approaches also agree well with each other and this is expected as the CNN emulates the manual segmentation approach, but does so on a much larger scale without any manual intervention. It must, however, be noted that a very small number of diffuse precipitates (close to the detection limit) remain unsegmented using the CNN approach as their grayscale values are hardly distinguishable from that of the matrix. It becomes impossible to track/detect the entirety of these small particles in 3D and can result in erroneous segmentation and are hence, omitted from segmentation. This is visible as a difference between the two approaches from the base of the histograms in Fig. 7a, in case of θ′ precipitates as they approach the detection limit.

[5] J.C.E. Mertens, J.J. Williams, N. Chawla, Development of a lab-scale, high-resolution, tube-generated X-ray computed-tomography system for three-dimensional (3D) materials characterization, Mater. Charact. 92 (2014) 36–48, http://dx.doi. org/10.1016/j.matchar.2014.03.002. [6] N. Limodin, J. Réthoré, J.Y. Buffière, A. Gravouil, F. Hild, S. Roux, Crack closure and stress intensity factor measurements in nodular graphite cast iron using threedimensional correlation of laboratory X-ray microtomography images, Acta Mater. 57 (2009) 4090–4101, http://dx.doi.org/10.1016/j.actamat.2009.05.005. [7] A.P. Merkle, J. Gelb, The ascent of 3D X-ray microscopy in the laboratory, Micros. Today 21 (2013) 10–15, http://dx.doi.org/10.1017/S1551929513000060. [8] H.A. Bale, A. Haboub, A.A. MacDowell, J.R. Nasiatka, D.Y. Parkinson, B.N. Cox, D.B. Marshall, R.O. Ritchie, Real-time quantitative imaging of failure events in materials under load at temperatures above 1,600 °C, Nat. Mater. 12 (2012) 40–46, http://dx.doi.org/10.1038/nmat3497. [9] N. Limodin, L. Salvo, E. Boller, M. Suéry, M. Felberbaum, S. Gailliègue, K. Madi, In situ and real-time 3-D microtomography investigation of dendritic solidification in an Al–10wt.% Cu alloy, Acta Mater. 57 (2009) 2300–2310, http://dx.doi.org/10. 1016/j.actamat.2009.01.035. [10] C.S. Kaira, V. De Andrade, S.S. Singh, C. Kantzos, A. Kirubanandham, F. De Carlo, N. Chawla, Probing novel microstructural evolution mechanisms in aluminum alloys using 4D nanoscale characterization, Adv. Mater. (2017) 1703482, , http://dx. doi.org/10.1002/adma.201703482. [11] S.W. Wilkins, T.E. Gureyev, D. Gao, A. Pogany, A.W. Stevenson, Phase-contrast imaging using polychromatic hard X-rays, Nature 384 (1996) 335–338, http://dx. doi.org/10.1038/384335a0. [12] J.A. Meganck, K.M. Kozloff, M.M. Thornton, S.M. Broski, S.A. Goldstein, Beam hardening artifacts in micro-computed tomography scanning can be reduced by Xray beam filtration and the resulting images can be used to accurately measure BMD, Bone 45 (2009) 1104–1116, http://dx.doi.org/10.1016/j.bone.2009.07.078. [13] O.B. Olurin, M. Arnold, C. Körner, R.F. Singer, The investigation of morphometric parameters of aluminium foams using micro-computed tomography, Mater. Sci. Eng. A 328 (2002) 334–343, http://dx.doi.org/10.1016/S0921-5093(01)01809-3. [14] E. Maire, P. Colombo, J. Adrien, L. Babout, L. Biasetto, Characterization of the morphology of cellular ceramics by 3D image processing of X-ray tomography, J. Eur. Ceram. Soc. 27 (2007) 1973–1981, http://dx.doi.org/10.1016/J. JEURCERAMSOC.2006.05.097. [15] C. Gupta, H. Toda, T. Fujioka, M. Kobayashi, H. Hoshino, K. Uesugi, A. Takeuchi, Y. Suzuki, State of 3-D micro-damage in hydrogen redistributed regions of precharged high strength aluminium alloy, Corros. Sci. 111 (2016) 26–38, http://dx. doi.org/10.1016/J.CORSCI.2016.04.050. [16] H. Toda, T. Kobayashi, M. Niinomi, T. Ohgaki, M. Kobayashi, N. Kuroda, T. Akahori, K. Uesugi, K. Makii, Y. Aruga, Quantitative assessment of microstructure and its effects on compression behavior of aluminum foams via high-resolution synchrotron X-ray tomography, Metall. Mater. Trans. A 37 (2006) 1211–1219, http://dx.doi.org/10.1007/s11661-006-1072-0. [17] H. Toda, I. Sinclair, J.-Y. Buffière, E. Maire, K.H. Khor, P. Gregson, T. Kobayashi, A 3D measurement procedure for internal local crack driving forces via synchrotron X-ray microtomography, Acta Mater. 52 (2004) 1305–1317, http://dx.doi.org/10. 1016/J.ACTAMAT.2003.11.014. [18] H. Toda, S. Yamamoto, M. Kobayashi, K. Uesugi, H. Zhang, Direct measurement procedure for three-dimensional local crack driving force using synchrotron X-ray microtomography, Acta Mater. 56 (2008) 6027–6039, http://dx.doi.org/10.1016/ J.ACTAMAT.2008.08.022. [19] E.N. Mortensen, W.A. Barrett, Interactive segmentation with intelligent scissors, Graph. Model. Image Process. 60 (1998) 349–384, http://dx.doi.org/10.1006/ gmip.1998.0480. [20] V. De Andrade, A. Deriy, M.J. Wojcik, D. Gürsoy, D. Shu, K. Fezzaa, F. De Carlo, Nanoscale 3D Imaging at the Advanced Photon Source, SPIE Newsroom, 2016, pp. 2–4, http://dx.doi.org/10.1117/2.1201604.006461. [21] S. Gorelick, J. Vila-Comamala, V.A. Guzenko, R. Barrett, M. Salomé, C. David, Highefficiency Fresnel zone plates for hard X-rays by 100 keV e-beam lithography and electroplating, J. Synchrotron Radiat. 18 (2011) 442–446, http://dx.doi.org/10. 1107/S0909049511002366. [22] H. Toda, K. Uesugi, A. Takeuchi, K. Minami, M. Kobayashi, T. Kobayashi, Threedimensional observation of nanoscopic precipitates in an aluminum alloy by microtomography with Fresnel zone plate optics, Appl. Phys. Lett. 89 (2006), http:// dx.doi.org/10.1063/1.2359288. [23] C.S. Kaira, C. Kantzos, J.J. Williams, V. De Andrade, F. De Carlo, N. Chawla, Microstructural evolution and deformation behavior of Al-Cu alloys: a Transmission X-ray Microscopy (TXM) and micropillar compression study, Acta Mater. 144 (2018) 419–431, http://dx.doi.org/10.1016/j.actamat.2017.11.009. [24] Y. LeCun, Y. Bengio, G. Hinton, L. Y, B. Y, H. G, Deep learning, Nature 521 (2015) 436–444, http://dx.doi.org/10.1038/nature14539. [25] S. Lawrence, C.L. Giles, Ah. Chung Tsoi, A.D. Back, Face recognition: a convolutional neural-network approach, IEEE Trans. Neural Netw. 8 (1997) 98–113, http:// dx.doi.org/10.1109/72.554195. [26] Y. LeCun, L. Bottou, Y. Bengio, P. Haffner, Gradient-based learning applied to document recognition, Proc. IEEE 86 (1998) 2278–2323, http://dx.doi.org/10. 1109/5.726791. [27] C. Cernazanu-Glavan, S. Holban, Segmentation of bone structure in X-ray images using convolutional neural network, Adv. Electr. Comput. Eng. 13 (2013) 87–94, http://dx.doi.org/10.4316/AECE.2013.01015. [28] X. Yang, F. De Carlo, C. Phatak, D. Gürsoy, A convolutional neural network approach to calibrating the rotation axis for X-ray computed tomography, J. Synchrotron Radiat. 24 (2017) 469–475, http://dx.doi.org/10.1107/ S1600577516020117.

4. Conclusions The effectiveness of utilizing a deep learning approach to perform automated segmentation using a Convolutional Neural Network (CNN) has been demonstrated on complex nanotomography datasets obtained using Transmission X-ray Microscopy (TXM). By coupling the same with Scanning Electron Microscopy imaging post-tomography and carefully selected image-processing filters, accurate quantification of the nanoscale precipitates' 3D morphology can be achieved. Quantitative comparison between the manual and CNN-based segmentation approaches validates the accuracy of the latter. This has enabled unprecedented swift and accurate quantification of tomographic data, with orders of magnitude improvement in the time scale required for analysis. This is crucial as it aims at removing one of the primary bottlenecks associated with 4D tomographic data analysis. Acknowledgements The authors are thankful for financial support from the Army Research Office (ARO) under Contract No. W911NF1410550 (Dr. Michael Bakas and Dr. David Stepp, Program Managers). We acknowledge the use of resources at Beamline 32-ID-C of the Advanced Photon Source, a U.S. Department of Energy (DOE) Office of Science User Facility operated for the DOE Office of Science by Argonne National Laboratory under Contract No. DE-AC02-06CH11357. We also acknowledge the use of facilities within the Center for 4DMS and the Leroy Eyring Center for Solid State Science at Arizona State University. References [1] E. Maire, P.J. Withers, Quantitative X-ray tomography, Int. Mater. Rev. 59 (2013) 1–43, http://dx.doi.org/10.1179/1743280413Y.0000000023. [2] J. Baruchel, J.-Y. Buffiere, P. Cloetens, M. Di Michiel, E. Ferrie, W. Ludwig, E. Maire, L. Salvo, Advances in synchrotron radiation microtomography, Scr. Mater. 55 (2006) 41–46, http://dx.doi.org/10.1016/j.scriptamat.2006.02.012. [3] B. Niemann, D. Rudolph, G. Schmahl, X-ray microscopy with synchrotron radiation, Appl. Opt. 15 (1976) 1883, http://dx.doi.org/10.1364/AO.15.001883. [4] J.J. Williams, Z. Flom, A.A. Amell, N. Chawla, X. Xiao, F. De Carlo, Damage evolution in SiC particle reinforced Al alloy matrix composites by X-ray synchrotron tomography, Acta Mater. 58 (2010) 6194–6205, http://dx.doi.org/10.1016/j. actamat.2010.07.039.

209



(1974) 957–969, http://dx.doi.org/10.1016/0001-6160(74)90021-2. [34] A. Guinier, Structure of age-hardened aluminium-copper alloys, Nature 142 (1938) 569–570, http://dx.doi.org/10.1038/142569b0. [35] G.D. Preston, Age-hardening of copper-aluminium alloys, Proc. Phys. Soc. 52 (1940) 77–79, http://dx.doi.org/10.1088/0959-5309/52/1/310. [36] C.A. Schneider, W.S. Rasband, K.W. Eliceiri, NIH image to ImageJ: 25 years of image analysis, Nat. Methods 9 (2012) 671–675, http://dx.doi.org/10.1038/nmeth. 2089. [37] A. Buades, B. Coll, J.-M.J.-M. Morel, A non-local algorithm for image denoising, Comput. Vis. Pattern Recognition, 2005. CVPR 2005, IEEE Comput. Soc. Conf. 2, 2 2005, pp. 60–65, , http://dx.doi.org/10.1109/CVPR.2005.38. [38] J. Canny, A computational approach to edge detection, IEEE Trans. Pattern Anal. Mach. Intell. 8 (1986) 679–698. [39] V. Vaithyanathan, C. Wolverton, L.Q. Chen, Multiscale modeling of θ′ precipitation in Al-Cu binary alloys, Acta Mater. 52 (2004) 2973–2987, http://dx.doi.org/10. 1016/j.actamat.2004.03.001.

[29] D. Gürsoy, F. De Carlo, X. Xiao, C. Jacobsen, TomoPy: a framework for the analysis of synchrotron tomographic data, J. Synchrotron Radiat. 21 (2014) 1188–1193, http://dx.doi.org/10.1107/S1600577514013939. [30] F. De Carlo, D. Gürsoy, F. Marone, M. Rivers, D.Y. Parkinson, F. Khan, N. Schwarz, D.J. Vine, S. Vogt, S.C. Gleber, S. Narayanan, M. Newville, T. Lanzirotti, Y. Sun, Y.P. Hong, C. Jacobsen, Scientific data exchange: a schema for HDF5-based storage of raw and analyzed data, J. Synchrotron Radiat. 21 (2014) 1224–1230, http://dx. doi.org/10.1107/S160057751401604X. [31] C. Laird, H. Aaronson, Mechanisms of formation of [theta] and dissolution of [theta] precipitates in an Al-4% Cu alloy, Acta Metall. 14 (1966) 171–185, http:// dx.doi.org/10.1016/0001-6160(66)90298-7. [32] J.F. Nie, B.C. Muddle, I.J. Polmear, The effect of precipitate shape and orientation on dispersion strengthening in high strength aluminium alloys, Mater. Sci. Forum 217–222 (1996) 1257–1262, http://dx.doi.org/10.4028/www.scientific.net/MSF. 217-222.1257. [33] R. Sankaran, C. Laird, Kinetics of growth of platelike precipitates, Acta Metall. 22

210